import requests import tensorflow as tf import pandas as pd import numpy as np from operator import add from functools import reduce import random import tabulate from keras import Model from keras import regularizers from keras.optimizers import Adam from keras.layers import Conv2D, BatchNormalization, ReLU, Input, Flatten, Softmax from keras.layers import Concatenate, Activation, Dense, GlobalAveragePooling2D, Dropout from keras.layers import AveragePooling1D, Bidirectional, LSTM, GlobalAveragePooling1D, MaxPool1D, Reshape from keras.layers import LayerNormalization, Conv1D, MultiHeadAttention, Layer from keras.models import load_model from keras.callbacks import EarlyStopping, ReduceLROnPlateau from Bio import SeqIO from Bio.SeqRecord import SeqRecord from Bio.SeqFeature import SeqFeature, FeatureLocation from Bio.Seq import Seq import cyvcf2 import parasail import re ntmap = {'A': (1, 0, 0, 0), 'C': (0, 1, 0, 0), 'G': (0, 0, 1, 0), 'T': (0, 0, 0, 1) } def get_seqcode(seq): return np.array(reduce(add, map(lambda c: ntmap[c], seq.upper()))).reshape((1, len(seq), -1)) class PositionalEncoding(Layer): def __init__(self, sequence_len=None, embedding_dim=None,**kwargs): super(PositionalEncoding, self).__init__() self.sequence_len = sequence_len self.embedding_dim = embedding_dim def call(self, x): position_embedding = np.array([ [pos / np.power(10000, 2. * i / self.embedding_dim) for i in range(self.embedding_dim)] for pos in range(self.sequence_len)]) position_embedding[:, 0::2] = np.sin(position_embedding[:, 0::2]) # dim 2i position_embedding[:, 1::2] = np.cos(position_embedding[:, 1::2]) # dim 2i+1 position_embedding = tf.cast(position_embedding, dtype=tf.float32) return position_embedding+x def get_config(self): config = super().get_config().copy() config.update({ 'sequence_len' : self.sequence_len, 'embedding_dim' : self.embedding_dim, }) return config def MultiHeadAttention_model(input_shape): input = Input(shape=input_shape) conv1 = Conv1D(256, 3, activation="relu")(input) pool1 = AveragePooling1D(2)(conv1) drop1 = Dropout(0.4)(pool1) conv2 = Conv1D(256, 3, activation="relu")(drop1) pool2 = AveragePooling1D(2)(conv2) drop2 = Dropout(0.4)(pool2) lstm = Bidirectional(LSTM(128, dropout=0.5, activation='tanh', return_sequences=True, kernel_regularizer=regularizers.l2(0.01)))(drop2) pos_embedding = PositionalEncoding(sequence_len=int(((23-3+1)/2-3+1)/2), embedding_dim=2*128)(lstm) atten = MultiHeadAttention(num_heads=2, key_dim=64, dropout=0.2, kernel_regularizer=regularizers.l2(0.01))(pos_embedding, pos_embedding) flat = Flatten()(atten) dense1 = Dense(512, kernel_regularizer=regularizers.l2(1e-4), bias_regularizer=regularizers.l2(1e-4), activation="relu")(flat) drop3 = Dropout(0.1)(dense1) dense2 = Dense(128, kernel_regularizer=regularizers.l2(1e-4), bias_regularizer=regularizers.l2(1e-4), activation="relu")(drop3) drop4 = Dropout(0.1)(dense2) dense3 = Dense(256, kernel_regularizer=regularizers.l2(1e-4), bias_regularizer=regularizers.l2(1e-4), activation="relu")(drop4) drop5 = Dropout(0.1)(dense3) output = Dense(1, activation="linear")(drop5) model = Model(inputs=[input], outputs=[output]) return model def fetch_ensembl_transcripts(gene_symbol): url = f"https://rest.ensembl.org/lookup/symbol/homo_sapiens/{gene_symbol}?expand=1;content-type=application/json" response = requests.get(url) if response.status_code == 200: gene_data = response.json() if 'Transcript' in gene_data: return gene_data['Transcript'] else: print("No transcripts found for gene:", gene_symbol) return None else: print(f"Error fetching gene data from Ensembl: {response.text}") return None def fetch_ensembl_sequence(transcript_id): url = f"https://rest.ensembl.org/sequence/id/{transcript_id}?content-type=application/json" response = requests.get(url) if response.status_code == 200: sequence_data = response.json() if 'seq' in sequence_data: return sequence_data['seq'] else: print("No sequence found for transcript:", transcript_id) return None else: print(f"Error fetching sequence data from Ensembl: {response.text}") return None def apply_mutation(ref_sequence, offset, ref, alt): """ Apply a single mutation to the sequence. """ if len(ref) == len(alt) and alt != "*": # SNP mutated_seq = ref_sequence[:offset] + alt + ref_sequence[offset+len(alt):] elif len(ref) < len(alt): # Insertion mutated_seq = ref_sequence[:offset] + alt + ref_sequence[offset+1:] elif len(ref) == len(alt) and alt == "*": # Deletion mutated_seq = ref_sequence[:offset] + ref_sequence[offset+1:] elif len(ref) > len(alt) and alt != "*": # Deletion mutated_seq = ref_sequence[:offset] + alt + ref_sequence[offset+len(ref):] elif len(ref) > len(alt) and alt == "*": # Deletion mutated_seq = ref_sequence[:offset] + ref_sequence[offset+len(ref):] return mutated_seq def construct_combinations(sequence, mutations): """ Construct all combinations of mutations. mutations is a list of tuples (position, ref, [alts]) """ if not mutations: return [sequence] # Take the first mutation and recursively construct combinations for the rest first_mutation = mutations[0] rest_mutations = mutations[1:] offset, ref, alts = first_mutation sequences = [] for alt in alts: mutated_sequence = apply_mutation(sequence, offset, ref, alt) sequences.extend(construct_combinations(mutated_sequence, rest_mutations)) return sequences def needleman_wunsch_alignment(query_seq, ref_seq): """ Use Needleman-Wunsch alignment to find the maximum alignment position in ref_seq Use this position to represent the position of target sequence with mutations """ # Needleman-Wunsch alignment alignment = parasail.nw_trace(query_seq, ref_seq, 10, 1, parasail.blosum62) # extract CIGAR object cigar = alignment.cigar cigar_string = cigar.decode.decode("utf-8") # record ref_pos ref_pos = 0 matches = re.findall(r'(\d+)([MIDNSHP=X])', cigar_string) max_num_before_equal = 0 max_equal_index = -1 total_before_max_equal = 0 for i, (num_str, op) in enumerate(matches): num = int(num_str) if op == '=': if num > max_num_before_equal: max_num_before_equal = num max_equal_index = i total_before_max_equal = sum(int(matches[j][0]) for j in range(max_equal_index)) ref_pos = total_before_max_equal return ref_pos def find_gRNA_with_mutation(ref_sequence, exon_chr, start, end, strand, transcript_id, exon_id, gene_symbol, vcf_reader, pam="NGG", target_length=20): # initialization mutated_sequences = [ref_sequence] # find mutations within interested region mutations = vcf_reader(f"{exon_chr}:{start}-{end}") if mutations: # find mutations mutation_list = [] for mutation in mutations: offset = mutation.POS - start ref = mutation.REF alts = mutation.ALT[:-1] mutation_list.append((offset, ref, alts)) # replace reference sequence of mutation mutated_sequences = construct_combinations(ref_sequence, mutation_list) # find gRNA in ref_sequence or all mutated_sequences targets = [] for seq in mutated_sequences: len_sequence = len(seq) dnatorna = {'A': 'A', 'T': 'U', 'C': 'C', 'G': 'G'} for i in range(len_sequence - len(pam) + 1): if seq[i + 1:i + 3] == pam[1:]: if i >= target_length: target_seq = seq[i - target_length:i + 3] pos = ref_sequence.find(target_seq) if pos != -1: is_mut = False if strand == -1: tar_start = end - pos - target_length - 2 else: tar_start = start + pos else: is_mut = True nw_pos = needleman_wunsch_alignment(target_seq, ref_sequence) if strand == -1: tar_start = str(end - nw_pos - target_length - 2) + '*' else: tar_start = str(start + nw_pos) + '*' gRNA = ''.join([dnatorna[base] for base in seq[i - target_length:i]]) targets.append([target_seq, gRNA, exon_chr, str(strand), str(tar_start), transcript_id, exon_id, gene_symbol, is_mut]) # filter duplicated targets unique_targets_set = set(tuple(element) for element in targets) unique_targets = [list(element) for element in unique_targets_set] return unique_targets def format_prediction_output_with_mutation(targets, model_path): model = MultiHeadAttention_model(input_shape=(23, 4)) model.load_weights(model_path) formatted_data = [] for target in targets: # Encode the gRNA sequence encoded_seq = get_seqcode(target[0]) # Predict on-target efficiency using the model prediction = float(list(model.predict(encoded_seq, verbose=0)[0])[0]) if prediction > 100: prediction = 100 # Format output gRNA = target[1] exon_chr = target[2] strand = target[3] tar_start = target[4] transcript_id = target[5] exon_id = target[6] gene_symbol = target[7] is_mut = target[8] formatted_data.append([gene_symbol, exon_chr, strand, tar_start, transcript_id, exon_id, target[0], gRNA, prediction, is_mut]) return formatted_data def process_gene(gene_symbol, vcf_reader, model_path): transcripts = fetch_ensembl_transcripts(gene_symbol) results = [] all_exons = [] # To accumulate all exons all_gene_sequences = [] # To accumulate all gene sequences if transcripts: for transcript in transcripts: Exons = transcript['Exon'] all_exons.extend(Exons) # Add all exons from this transcript to the list transcript_id = transcript['id'] for Exon in Exons: exon_id = Exon['id'] gene_sequence = fetch_ensembl_sequence(exon_id) # Reference exon sequence if gene_sequence: all_gene_sequences.append(gene_sequence) # Add this gene sequence to the list exon_chr = Exon['seq_region_name'] start = Exon['start'] end = Exon['end'] strand = Exon['strand'] targets = find_gRNA_with_mutation(gene_sequence, exon_chr, start, end, strand, transcript_id, exon_id, gene_symbol, vcf_reader) if targets: # Predict on-target efficiency for each gRNA site including mutations formatted_data = format_prediction_output_with_mutation(targets, model_path) results.extend(formatted_data) else: print(f"Failed to retrieve gene sequence for exon {exon_id}.") else: print("Failed to retrieve transcripts.") # Return the sorted output, combined gene sequences, and all exons return results, all_gene_sequences, all_exons def create_genbank_features(data): features = [] # If the input data is a DataFrame, convert it to a list of lists if isinstance(data, pd.DataFrame): formatted_data = data.values.tolist() elif isinstance(data, list): formatted_data = data else: raise TypeError("Data should be either a list or a pandas DataFrame.") for row in formatted_data: try: start = int(row[1]) end = start + len(row[6]) # Calculate the end position based on the target sequence length except ValueError as e: print(f"Error converting start/end to int: {row[1]}, {row[2]} - {e}") continue strand = 1 if row[3] == '1' else -1 location = FeatureLocation(start=start, end=end, strand=strand) is_mutation = 'Yes' if row[9] else 'No' feature = SeqFeature(location=location, type="misc_feature", qualifiers={ 'label': row[7], # Use gRNA as the label 'note': f"Prediction: {row[8]}, Mutation: {is_mutation}" # Include the prediction score and mutation status }) features.append(feature) return features def generate_genbank_file_from_df(df, gene_sequence, gene_symbol, output_path): # Ensure gene_sequence is a string before creating Seq object if not isinstance(gene_sequence, str): gene_sequence = str(gene_sequence) features = create_genbank_features(df) # Now gene_sequence is guaranteed to be a string, suitable for Seq seq_obj = Seq(gene_sequence) record = SeqRecord(seq_obj, id=gene_symbol, name=gene_symbol, description=f'CRISPR Cas9 predicted targets for {gene_symbol}', features=features) record.annotations["molecule_type"] = "DNA" SeqIO.write(record, output_path, "genbank") def create_bed_file_from_df(df, output_path): with open(output_path, 'w') as bed_file: for index, row in df.iterrows(): chrom = row["Chr"] start = int(row["Target Start"]) end = start + len(row["Target"]) # Calculate the end position based on the target sequence length strand = '+' if row["Strand"] == '1' else '-' gRNA = row["gRNA"] score = str(row["Prediction"]) is_mutation = 'Yes' if row["Is Mutation"] else 'No' # transcript_id is not typically part of the standard BED columns but added here for completeness transcript_id = row["Transcript"] # Writing only standard BED columns; additional columns can be appended as needed bed_file.write(f"{chrom}\t{start}\t{end}\t{gRNA}\t{score}\t{strand}\t{is_mutation}\n") def create_csv_from_df(df, output_path): df.to_csv(output_path, index=False)